Controlling Procedural Modeling Programs with Stochastically-Ordered Sequential Monte Carlo

摘要

We present a method for controlling the output of proceduralrnmodeling programs using Sequential Monte Carlo (SMC). Previousrnprobabilistic methods for controlling procedural models usernMarkov Chain Monte Carlo (MCMC), which receives control feedbackrnonly for completely-generated models. In contrast, SMC receivesrnfeedback incrementally on incomplete models, allowing it tornreallocate computational resources and converge quickly. To handlernthe many possible sequentializations of a structured, recursive proceduralrnmodeling program, we develop and prove the correctnessrnof a new SMC variant, Stochastically-Ordered Sequential MonternCarlo (SOSMC). We implement SOSMC for general-purpose programsrnusing a new programming primitive: the stochastic future.rnFinally, we show that SOSMC reliably generates high-quality outputsrnfor a variety of programs and control scoring functions. Forrnsmall computational budgets, SOSMC’s outputs often score nearlyrntwice as high as those of MCMC or normal SMC.